{
  "id": 16970092,
  "title": "Why ROUGE Lies About Your LLM Outlines and How Grounded Fine-Tuning Fixes It",
  "url": "https://urgent.news/2026/10/28/why-rouge-lies-about-your-llm-outlines-and-how-grounded-fine-tuning",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-28T21:41:19.000Z",
  "source": {
    "name": "HackerNoon",
    "slug": "hackernoon",
    "url": "https://hackernoon.com/why-rouge-lies-about-your-llm-outlines-and-how-grounded-fine-tuning-fixes-it?source=rss"
  },
  "original_language": "en",
  "account": "Existing approaches to outline generation for large language models (LLMs) face challenges in accurately creating structured outputs. This article proposes a grounded fine-tuning (GFT) method that addresses these issues by using paragraph IDs as a common basis for labeling, model output, and evaluation. The GFT pipeline consists of several key components:\n\n1. Paragraph segmentation: Paragraphs are segmented into sections using a supervised learning approach, assigning each paragraph a unique ID (e.g., [p17]) that remains constant throughout the document. This approach overcomes the limitations of free-form headings and enables the model to understand the document's structure.\n\n2. Hybrid label builder: The pipeline first identifies paragraph boundaries, then generates titles for each section based on either the existing heading or the best lead sentence. The titles are created by converting generative phrases into short, paraphrased versions. This hybrid approach combines the accuracy of extractive methods with the ability of abstractive methods to generate coherent content.\n\n3. Evaluation and assessment: The GFT pipeline includes multiple assessors that connect evaluation results to the human reader's judgment. These assessors compare the generated outlines to extractive-only and generative-only labeling methods on the same held-out dataset. The pipeline also reports the percentage of documents with acceptable outlines and compares this to the human-audited share.\n\n4. Addressing noisy span boundaries: The GFT pipeline includes data-quality checks to ensure accurate span boundaries, preventing confabulation and improving the reliability of the generated outlines.\n\nBy implementing this GFT approach, the outlined method aims to produce more accurate and reliable structured outputs from LLMs, ultimately improving the usability and effectiveness of such models in various applications.",
  "summary": "Fine-tune LLMs to generate document outlines tied to paragraph IDs, using hybrid extractive–generative labels and layered evaluators instead of ROUGE.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}